Linguistic Markers of Cognition

Semantic association instruments rest on a fundamental assumption: that the words people produce reveal something stable about how their minds organize meaning. If this assumption fails—if word choice reflects only communication style, momentary context, or measurement noise—then INS-001 measures nothing coherent.

This library article establishes the theoretical warrant for treating linguistic output as a window into cognition. Three claims must hold:

1. Semantic networks vary meaningfully across individuals (H1.2) — People differ in how concepts connect in memory, and these differences predict creative behavior
2. Embedding distances approximate cognitive semantic distance (H1.7) — Computational measures of word similarity track psychological reality
3. Semantic spread correlates with creativity (H1.8) — The Divergent Association Task methodology captures something real about divergent thinking

Each claim carries caveats documented below and empirically examined in MTH-002.1.

1. Why This Matters for INS-001

Semantic association instruments rest on a fundamental assumption: that the words people produce reveal something stable about how their minds organize meaning. If this assumption fails—if word choice reflects only communication style, momentary context, or measurement noise—then INS-001 measures nothing coherent.

This library article establishes the theoretical warrant for treating linguistic output as a window into cognition. Three claims must hold:

1. Semantic networks vary meaningfully across individuals (H1.2) — People differ in how concepts connect in memory, and these differences predict creative behavior
2. Embedding distances approximate cognitive semantic distance (H1.7) — Computational measures of word similarity track psychological reality
3. Semantic spread correlates with creativity (H1.8) — The Divergent Association Task methodology captures something real about divergent thinking

Each claim carries caveats documented below and empirically examined in MTH-002.1.

2. Enables

3. Must Establish

H1.2: Semantic Networks Vary Meaningfully

Claim: Individuals differ in the topology of their semantic networks—how densely concepts cluster, how efficiently paths connect distant ideas—and these structural differences predict creative performance.

Status: Supported

Evidence:

• Kenett et al. (2014) demonstrated that highly creative individuals show lower clustering and shorter path lengths in semantic networks derived from free associations
• Beaty & Kenett (2023) synthesized associative theories of creativity, establishing that network structure reflects individual differences in creative cognition

Implication for INS-001: We can meaningfully compare how individuals traverse semantic space, not merely record random variation.

H1.7: Embedding Validity for Semantic Distance

Claim: Computational word embeddings (vector representations learned from text corpora) approximate the semantic distances humans perceive between concepts.

Status: Partially supported at population level; individual-level validity untested

Evidence:

• Hill et al. (2015) validated embeddings against SimLex-999, a benchmark of human similarity judgments. However, SimLex-999 measures aggregate human judgments, not individual differences.
• Auguste et al. (2017) showed embeddings predict semantic priming reaction times at population level, providing cognitive validation. Individual-difference validity was not tested.

Complication: MTH-002.1 §4.3 documents that different embedding models produce substantially different scores:

• GloVe vs. OpenAI text-embedding-3-small: r = 0.60 (moderate, not interchangeable)
• Systematic difference: 20.7 points on the DAT scale

Implication for INS-001: Spread scores are model-dependent. A score of 65 using OpenAI embeddings is not equivalent to 65 using GloVe. We standardize on OpenAI text-embedding-3-small, but this choice affects interpretation.

H1.8: Semantic Spread Correlates with Creativity

Claim: The Divergent Association Task—asking participants to generate 10 unrelated words and measuring mean pairwise semantic distance—captures divergent thinking ability.

Status: Supported for 10-word unconstrained task; transfer to brief constrained format empirically calibrated but not externally validated

Evidence:

• Olson et al. (2021) validated DAT against established creativity measures: r = 0.40 with Alternative Uses Task (AUT), r = 0.28 with Remote Associates Test (RAT)
• Said-Metwaly et al. (2024) meta-analysis of divergent thinking and creative achievement finds r = 0.18, calibrating effect size expectations

Complication: INS-001 uses 2–5 words under constraint rather than 10 words unconstrained. MTH-002.1 §3.3 documents this difference empirically:

• INS-001.1 spread: 8.6 points lower than DAT reference
• INS-001.2 spread: 8.0 points lower than DAT reference

The gap is documented but its meaning is ambiguous: does it reflect reduced construct validity, or simply geometric compression from constraints?

Implication for INS-001: Spread measures something related to the DAT construct, but the relationship is calibrated rather than validated.

4. Key Sources

5. Scope

In scope:

• Semantic network structure and creativity (theoretical basis)
• Embedding validity for semantic distance (scoring rationale)
• Divergent Association Task methodology (spread metric)
• Free vs. goal-directed association distinction (instrument design)

Out of scope:

• Platform effects on expression → LIB-003
• Digital measurement validity → LIB-002
• Game-based assessment → LIB-008
• Cross-cultural invariance → LIB-007 (Note: MTH-002.1 §9 explicitly scopes INS-001 to English only)

6. Current Gaps

7. Confidence

Low-to-Moderate.

H1.2 (semantic networks vary meaningfully) has strong empirical support from the Kenett lab. The theoretical framework is well-established.

H1.8 (DAT validity) is supported for the original 10-word task, but transfer to INS-001’s brief constrained format is empirically calibrated (MTH-002.1) without external criterion validation. We know INS-001 produces lower spread scores; we don’t know if this affects what the scores mean.

H1.7 (embedding validity) is documented for population semantics but shows large model-dependent effects (20.7 points) that complicate individual-level interpretation. Embeddings approximate group-level semantic judgments; whether they capture meaningful individual differences is an empirical question we have not answered.

The stability claim remains unvalidated—no test-retest data exists for semantic spread measures.

8. Joint Confidence Note

Individual confidence ratings treat gaps as independent. When multiple LIB articles are combined for instrument development (e.g., INS-001 depends on LIB-001, LIB-002, and LIB-008), joint confidence is substantially lower than any single rating suggests.

Key compounding uncertainties for INS-001:

• Embedding validity for individuals (LIB-001 H1.7) × Test-retest reliability (LIB-002 H2.3) × Creativity-game transfer (LIB-008 H8.1)

Until these are independently validated, INS-001 confidence should be interpreted as Low despite moderate ratings on component claims.

Changelog